An AMD engineer presented earlier this month at the Vulkanised 2025 conference in Cambridge (UK) around the work they are pursuing for AI using the MLIR intermediate representation, IREE, and the role that Vulkan/SPIR-V can play for AI acceleration across AMD’s wares as well as other hardware.
We’ve known for a long while that MLIR will be central to AMD’s elusive “Unified AI Software Stack” as a common IR to then target their different hardware from Radeon GPUs, Instinct accelerators, Ryzen AI NPUs, and CPU-based execution in a unified manner. They’ve also been working on a generic MLIR to SPIR-V target that in turn could potentially open up execution to non-AMD SPIR-V supported drivers.
As I wrote last summer, AMD’s Unified AI Software Stack Might Be A Boon For Other Vulkan/SPIR-V Hardware Too. In more recent months they have continued hiring around MLIR and IREE developer talent and talking up Ryzen AI programming with MLIR and IREE.
At the seventh annual Vulkanised conference earlier this month hosted by The Khronos Group, AMD engineer Jakub Kuderski was talking up AI programming with Vulkan’s SPIR-V IR in IREE and MLIR. For those interested in this AI programming model being pursued by AMD, the presentation video was uploaded to YouTube yesterday:
There is also the slide deck reaffirming this model and the positive results they are seeing, including on both Radeon and Instinct hardware.
Noted ongoing work is being able to support heterogeneous executables with GPU and NPU kernels within the same program binary, “one click” whole-model tuning using profile guided optimizations (PGO), and other functionality. Interesting work ahead and hopefully more of AMD’s unified AI software efforts come to fruition in 2025.
